为构建离子液体的CO2溶解度的准确预测模型,考虑到传统模型存在的描述符计算复杂、成本高、关联结构与性质困难、结构特征提取不充分等问题,提出一种融合了加入注意力机制的图卷积神经网络和XGBoost的预测模型(APGCN-XGBoost).对9 897 组离子液体的CO2溶解度数据的分析结果显示,所提出的APGCN-XGBoost模型在预测性能上优于传统的分子指纹模型和图卷积神经网络模型.此外,通过注意力池化层与SHAP方法对模型进行解释,APGCN-XGBoost模型学习到了离子液体中各个原子和结构的特征信息与分子非局部信息,这些特征信息不仅可以用于性质预测,还可以用于探索化学结构与性质之间的联系,即通过模型的解释,筛选出对于溶解度预测重要的离子液体结构信息,从而实现CO2捕获过程中理想离子液体的计算机辅助设计和筛选.
In order to build an accurate prediction model of CO2 solubility of ionic liquids,considering the problems existing in traditional models,such as complex descriptor calculation,high cost,difficult related structure and properties,and insufficient structural feature extraction,a prediction model APGCN-XGBoost combining graph convolution neural network and XGBoost with attention mechanism was proposed.The analysis results of CO2 solubility data of 9 897 groups of ionic liquids show that the prediction performance of the proposed APGCN-XGBoost model is better than that of the traditional molecular fingerprint model and graph convo-lutional neural network model.In addition,the APGCN-XGBoost model learned the characteristic information and molecular non-local information of each atom and structure in ionic liquids,which can be used not only to predict the properties,but also to explore the rela-tionship between the chemical structures and properties,that is,to screen out the structural information of ionic liquids that is important for solubility prediction through the interpretation of the model,thus realizing the computer-aided design and screening of ideal ionic liquids in the process of CO2 capture.